path: root/libpod/oci_conmon_linux.go

// +build linux

package libpod

import (
	"bufio"
	"bytes"
	"context"
	"fmt"
	"io"
	"io/ioutil"
	"net"
	"net/http"
	"os"
	"os/exec"
	"path/filepath"
	"runtime"
	"strconv"
	"strings"
	"sync"
	"syscall"
	"text/template"
	"time"

	"github.com/containers/common/pkg/capabilities"
	"github.com/containers/common/pkg/config"
	conmonConfig "github.com/containers/conmon/runner/config"
	"github.com/containers/podman/v3/libpod/define"
	"github.com/containers/podman/v3/libpod/logs"
	"github.com/containers/podman/v3/pkg/cgroups"
	"github.com/containers/podman/v3/pkg/errorhandling"
	"github.com/containers/podman/v3/pkg/lookup"
	"github.com/containers/podman/v3/pkg/rootless"
	"github.com/containers/podman/v3/pkg/util"
	"github.com/containers/podman/v3/utils"
	"github.com/containers/storage/pkg/homedir"
	pmount "github.com/containers/storage/pkg/mount"
	"github.com/coreos/go-systemd/v22/activation"
	"github.com/coreos/go-systemd/v22/daemon"
	spec "github.com/opencontainers/runtime-spec/specs-go"
	"github.com/opencontainers/selinux/go-selinux"
	"github.com/opencontainers/selinux/go-selinux/label"
	"github.com/pkg/errors"
	"github.com/sirupsen/logrus"
	"golang.org/x/sys/unix"
	"k8s.io/client-go/tools/remotecommand"
)

const (
	// This is Conmon's STDIO_BUF_SIZE. I don't believe we have access to it
	// directly from the Go code, so const it here
	bufferSize = conmonConfig.BufSize
)

// ConmonOCIRuntime is an OCI runtime managed by Conmon.
// TODO: Make all calls to OCI runtime have a timeout.
type ConmonOCIRuntime struct {
	name              string
	path              string
	conmonPath        string
	conmonEnv         []string
	tmpDir            string
	exitsDir          string
	socketsDir        string
	logSizeMax        int64
	noPivot           bool
	reservePorts      bool
	runtimeFlags      []string
	supportsJSON      bool
	supportsKVM       bool
	supportsNoCgroups bool
	sdNotify          bool
	enableKeyring     bool
}

// Make a new Conmon-based OCI runtime with the given options.
// Conmon will wrap the given OCI runtime, which can be `runc`, `crun`, or
// any runtime with a runc-compatible CLI.
// The first path that points to a valid executable will be used.
// Deliberately private. Someone should not be able to construct this outside of
// libpod.
func newConmonOCIRuntime(name string, paths []string, conmonPath string, runtimeFlags []string, runtimeCfg *config.Config) (OCIRuntime, error) {
	if name == "" {
		return nil, errors.Wrapf(define.ErrInvalidArg, "the OCI runtime must be provided a non-empty name")
	}

	// Make lookup tables for runtime support
	supportsJSON := make(map[string]bool, len(runtimeCfg.Engine.RuntimeSupportsJSON))
	supportsNoCgroups := make(map[string]bool, len(runtimeCfg.Engine.RuntimeSupportsNoCgroups))
	supportsKVM := make(map[string]bool, len(runtimeCfg.Engine.RuntimeSupportsKVM))
	for _, r := range runtimeCfg.Engine.RuntimeSupportsJSON {
		supportsJSON[r] = true
	}
	for _, r := range runtimeCfg.Engine.RuntimeSupportsNoCgroups {
		supportsNoCgroups[r] = true
	}
	for _, r := range runtimeCfg.Engine.RuntimeSupportsKVM {
		supportsKVM[r] = true
	}

	runtime := new(ConmonOCIRuntime)
	runtime.name = name
	runtime.conmonPath = conmonPath
	runtime.runtimeFlags = runtimeFlags

	runtime.conmonEnv = runtimeCfg.Engine.ConmonEnvVars
	runtime.tmpDir = runtimeCfg.Engine.TmpDir
	runtime.logSizeMax = runtimeCfg.Containers.LogSizeMax
	runtime.noPivot = runtimeCfg.Engine.NoPivotRoot
	runtime.reservePorts = runtimeCfg.Engine.EnablePortReservation
	runtime.sdNotify = runtimeCfg.Engine.SDNotify
	runtime.enableKeyring = runtimeCfg.Containers.EnableKeyring

	// TODO: probe OCI runtime for feature and enable automatically if
	// available.
	runtime.supportsJSON = supportsJSON[name]
	runtime.supportsNoCgroups = supportsNoCgroups[name]
	runtime.supportsKVM = supportsKVM[name]

	foundPath := false
	for _, path := range paths {
		stat, err := os.Stat(path)
		if err != nil {
			if os.IsNotExist(err) {
				continue
			}
			return nil, errors.Wrapf(err, "cannot stat OCI runtime %s path", name)
		}
		if !stat.Mode().IsRegular() {
			continue
		}
		foundPath = true
		runtime.path = path
		logrus.Debugf("using runtime %q", path)
		break
	}

	// Search the $PATH as last fallback
	if !foundPath {
		if foundRuntime, err := exec.LookPath(name); err == nil {
			foundPath = true
			runtime.path = foundRuntime
			logrus.Debugf("using runtime %q from $PATH: %q", name, foundRuntime)
		}
	}

	if !foundPath {
		return nil, errors.Wrapf(define.ErrInvalidArg, "no valid executable found for OCI runtime %s", name)
	}

	runtime.exitsDir = filepath.Join(runtime.tmpDir, "exits")
	runtime.socketsDir = filepath.Join(runtime.tmpDir, "socket")

	// Create the exit files and attach sockets directories
	if err := os.MkdirAll(runtime.exitsDir, 0750); err != nil {
		// The directory is allowed to exist
		if !os.IsExist(err) {
			return nil, errors.Wrapf(err, "error creating OCI runtime exit files directory")
		}
	}
	if err := os.MkdirAll(runtime.socketsDir, 0750); err != nil {
		// The directory is allowed to exist
		if !os.IsExist(err) {
			return nil, errors.Wrap(err, "error creating OCI runtime attach sockets directory")
		}
	}

	return runtime, nil
}

// Name returns the name of the runtime being wrapped by Conmon.
func (r *ConmonOCIRuntime) Name() string {
	return r.name
}

// Path returns the path of the OCI runtime being wrapped by Conmon.
func (r *ConmonOCIRuntime) Path() string {
	return r.path
}

// hasCurrentUserMapped checks whether the current user is mapped inside the container user namespace
func hasCurrentUserMapped(ctr *Container) bool {
	if len(ctr.config.IDMappings.UIDMap) == 0 && len(ctr.config.IDMappings.GIDMap) == 0 {
		return true
	}
	uid := os.Geteuid()
	for _, m := range ctr.config.IDMappings.UIDMap {
		if uid >= m.HostID && uid < m.HostID+m.Size {
			return true
		}
	}
	return false
}

// CreateContainer creates a container.
func (r *ConmonOCIRuntime) CreateContainer(ctr *Container, restoreOptions *ContainerCheckpointOptions) error {
	// always make the run dir accessible to the current user so that the PID files can be read without
	// being in the rootless user namespace.
	if err := makeAccessible(ctr.state.RunDir, 0, 0); err != nil {
		return err
	}
	if !hasCurrentUserMapped(ctr) {
		for _, i := range []string{ctr.state.RunDir, ctr.runtime.config.Engine.TmpDir, ctr.config.StaticDir, ctr.state.Mountpoint, ctr.runtime.config.Engine.VolumePath} {
			if err := makeAccessible(i, ctr.RootUID(), ctr.RootGID()); err != nil {
				return err
			}
		}

		// if we are running a non privileged container, be sure to umount some kernel paths so they are not
		// bind mounted inside the container at all.
		if !ctr.config.Privileged && !rootless.IsRootless() {
			ch := make(chan error)
			go func() {
				runtime.LockOSThread()
				err := func() error {
					fd, err := os.Open(fmt.Sprintf("/proc/%d/task/%d/ns/mnt", os.Getpid(), unix.Gettid()))
					if err != nil {
						return err
					}
					defer errorhandling.CloseQuiet(fd)

					// create a new mountns on the current thread
					if err = unix.Unshare(unix.CLONE_NEWNS); err != nil {
						return err
					}
					defer func() {
						if err := unix.Setns(int(fd.Fd()), unix.CLONE_NEWNS); err != nil {
							logrus.Errorf("unable to clone new namespace: %q", err)
						}
					}()

					// don't spread our mounts around.  We are setting only /sys to be slave
					// so that the cleanup process is still able to umount the storage and the
					// changes are propagated to the host.
					err = unix.Mount("/sys", "/sys", "none", unix.MS_REC|unix.MS_SLAVE, "")
					if err != nil {
						return errors.Wrapf(err, "cannot make /sys slave")
					}

					mounts, err := pmount.GetMounts()
					if err != nil {
						return err
					}
					for _, m := range mounts {
						if !strings.HasPrefix(m.Mountpoint, "/sys/kernel") {
							continue
						}
						err = unix.Unmount(m.Mountpoint, 0)
						if err != nil && !os.IsNotExist(err) {
							return errors.Wrapf(err, "cannot unmount %s", m.Mountpoint)
						}
					}
					return r.createOCIContainer(ctr, restoreOptions)
				}()
				ch <- err
			}()
			err := <-ch
			return err
		}
	}
	return r.createOCIContainer(ctr, restoreOptions)
}

// UpdateContainerStatus retrieves the current status of the container from the
// runtime. It updates the container's state but does not save it.
// If useRuntime is false, we will not directly hit runc to see the container's
// status, but will instead only check for the existence of the conmon exit file
// and update state to stopped if it exists.
func (r *ConmonOCIRuntime) UpdateContainerStatus(ctr *Container) error {
	exitFile, err := r.ExitFilePath(ctr)
	if err != nil {
		return err
	}

	runtimeDir, err := util.GetRuntimeDir()
	if err != nil {
		return err
	}

	// Store old state so we know if we were already stopped
	oldState := ctr.state.State

	state := new(spec.State)

	cmd := exec.Command(r.path, "state", ctr.ID())
	cmd.Env = append(cmd.Env, fmt.Sprintf("XDG_RUNTIME_DIR=%s", runtimeDir))

	outPipe, err := cmd.StdoutPipe()
	if err != nil {
		return errors.Wrapf(err, "getting stdout pipe")
	}
	errPipe, err := cmd.StderrPipe()
	if err != nil {
		return errors.Wrapf(err, "getting stderr pipe")
	}

	if err := cmd.Start(); err != nil {
		out, err2 := ioutil.ReadAll(errPipe)
		if err2 != nil {
			return errors.Wrapf(err, "error getting container %s state", ctr.ID())
		}
		if strings.Contains(string(out), "does not exist") {
			if err := ctr.removeConmonFiles(); err != nil {
				logrus.Debugf("unable to remove conmon files for container %s", ctr.ID())
			}
			ctr.state.ExitCode = -1
			ctr.state.FinishedTime = time.Now()
			ctr.state.State = define.ContainerStateExited
			return nil
		}
		return errors.Wrapf(err, "error getting container %s state. stderr/out: %s", ctr.ID(), out)
	}
	defer func() {
		_ = cmd.Wait()
	}()

	if err := errPipe.Close(); err != nil {
		return err
	}
	out, err := ioutil.ReadAll(outPipe)
	if err != nil {
		return errors.Wrapf(err, "error reading stdout: %s", ctr.ID())
	}
	if err := json.NewDecoder(bytes.NewBuffer(out)).Decode(state); err != nil {
		return errors.Wrapf(err, "error decoding container status for container %s", ctr.ID())
	}
	ctr.state.PID = state.Pid

	switch state.Status {
	case "created":
		ctr.state.State = define.ContainerStateCreated
	case "paused":
		ctr.state.State = define.ContainerStatePaused
	case "running":
		ctr.state.State = define.ContainerStateRunning
	case "stopped":
		ctr.state.State = define.ContainerStateStopped
	default:
		return errors.Wrapf(define.ErrInternal, "unrecognized status returned by runtime for container %s: %s",
			ctr.ID(), state.Status)
	}

	// Only grab exit status if we were not already stopped
	// If we were, it should already be in the database
	if ctr.state.State == define.ContainerStateStopped && oldState != define.ContainerStateStopped {
		var fi os.FileInfo
		chWait := make(chan error)
		defer close(chWait)

		_, err := WaitForFile(exitFile, chWait, time.Second*5)
		if err == nil {
			fi, err = os.Stat(exitFile)
		}
		if err != nil {
			ctr.state.ExitCode = -1
			ctr.state.FinishedTime = time.Now()
			logrus.Errorf("No exit file for container %s found: %v", ctr.ID(), err)
			return nil
		}

		return ctr.handleExitFile(exitFile, fi)
	}

	return nil
}

// StartContainer starts the given container.
// Sets time the container was started, but does not save it.
func (r *ConmonOCIRuntime) StartContainer(ctr *Container) error {
	// TODO: streams should probably *not* be our STDIN/OUT/ERR - redirect to buffers?
	runtimeDir, err := util.GetRuntimeDir()
	if err != nil {
		return err
	}
	env := []string{fmt.Sprintf("XDG_RUNTIME_DIR=%s", runtimeDir)}
	if ctr.config.SdNotifyMode == define.SdNotifyModeContainer {
		if notify, ok := os.LookupEnv("NOTIFY_SOCKET"); ok {
			env = append(env, fmt.Sprintf("NOTIFY_SOCKET=%s", notify))
		}
	}
	if path, ok := os.LookupEnv("PATH"); ok {
		env = append(env, fmt.Sprintf("PATH=%s", path))
	}
	if err := utils.ExecCmdWithStdStreams(os.Stdin, os.Stdout, os.Stderr, env, r.path, append(r.runtimeFlags, "start", ctr.ID())...); err != nil {
		return err
	}

	ctr.state.StartedTime = time.Now()

	return nil
}

// KillContainer sends the given signal to the given container.
// If all is set, send to all PIDs in the container.
// All is only supported if the container created cgroups.
func (r *ConmonOCIRuntime) KillContainer(ctr *Container, signal uint, all bool) error {
	logrus.Debugf("Sending signal %d to container %s", signal, ctr.ID())
	runtimeDir, err := util.GetRuntimeDir()
	if err != nil {
		return err
	}
	env := []string{fmt.Sprintf("XDG_RUNTIME_DIR=%s", runtimeDir)}
	var args []string
	args = append(args, r.runtimeFlags...)
	if all {
		args = append(args, "kill", "--all", ctr.ID(), fmt.Sprintf("%d", signal))
	} else {
		args = append(args, "kill", ctr.ID(), fmt.Sprintf("%d", signal))
	}
	if err := utils.ExecCmdWithStdStreams(os.Stdin, os.Stdout, os.Stderr, env, r.path, args...); err != nil {
		return errors.Wrapf(err, "error sending signal to container %s", ctr.ID())
	}

	return nil
}

// StopContainer stops a container, first using its given stop signal (or
// SIGTERM if no signal was specified), then using SIGKILL.
// Timeout is given in seconds. If timeout is 0, the container will be
// immediately kill with SIGKILL.
// Does not set finished time for container, assumes you will run updateStatus
// after to pull the exit code.
func (r *ConmonOCIRuntime) StopContainer(ctr *Container, timeout uint, all bool) error {
	logrus.Debugf("Stopping container %s (PID %d)", ctr.ID(), ctr.state.PID)

	// Ping the container to see if it's alive
	// If it's not, it's already stopped, return
	err := unix.Kill(ctr.state.PID, 0)
	if err == unix.ESRCH {
		return nil
	}

	stopSignal := ctr.config.StopSignal
	if stopSignal == 0 {
		stopSignal = uint(syscall.SIGTERM)
	}

	if timeout > 0 {
		if err := r.KillContainer(ctr, stopSignal, all); err != nil {
			// Is the container gone?
			// If so, it probably died between the first check and
			// our sending the signal
			// The container is stopped, so exit cleanly
			err := unix.Kill(ctr.state.PID, 0)
			if err == unix.ESRCH {
				return nil
			}

			return err
		}

		if err := waitContainerStop(ctr, time.Duration(timeout)*time.Second); err != nil {
			logrus.Infof("Timed out stopping container %s, resorting to SIGKILL: %v", ctr.ID(), err)
		} else {
			// No error, the container is dead
			return nil
		}
	}

	if err := r.KillContainer(ctr, 9, all); err != nil {
		// Again, check if the container is gone. If it is, exit cleanly.
		err := unix.Kill(ctr.state.PID, 0)
		if err == unix.ESRCH {
			return nil
		}

		return errors.Wrapf(err, "error sending SIGKILL to container %s", ctr.ID())
	}

	// Give runtime a few seconds to make it happen
	if err := waitContainerStop(ctr, killContainerTimeout); err != nil {
		return err
	}

	return nil
}

// DeleteContainer deletes a container from the OCI runtime.
func (r *ConmonOCIRuntime) DeleteContainer(ctr *Container) error {
	runtimeDir, err := util.GetRuntimeDir()
	if err != nil {
		return err
	}
	env := []string{fmt.Sprintf("XDG_RUNTIME_DIR=%s", runtimeDir)}
	return utils.ExecCmdWithStdStreams(os.Stdin, os.Stdout, os.Stderr, env, r.path, append(r.runtimeFlags, "delete", "--force", ctr.ID())...)
}

// PauseContainer pauses the given container.
func (r *ConmonOCIRuntime) PauseContainer(ctr *Container) error {
	runtimeDir, err := util.GetRuntimeDir()
	if err != nil {
		return err
	}
	env := []string{fmt.Sprintf("XDG_RUNTIME_DIR=%s", runtimeDir)}
	return utils.ExecCmdWithStdStreams(os.Stdin, os.Stdout, os.Stderr, env, r.path, append(r.runtimeFlags, "pause", ctr.ID())...)
}

// UnpauseContainer unpauses the given container.
func (r *ConmonOCIRuntime) UnpauseContainer(ctr *Container) error {
	runtimeDir, err := util.GetRuntimeDir()
	if err != nil {
		return err
	}
	env := []string{fmt.Sprintf("XDG_RUNTIME_DIR=%s", runtimeDir)}
	return utils.ExecCmdWithStdStreams(os.Stdin, os.Stdout, os.Stderr, env, r.path, append(r.runtimeFlags, "resume", ctr.ID())...)
}

// HTTPAttach performs an attach for the HTTP API.
// The caller must handle closing the HTTP connection after this returns.
// The cancel channel is not closed; it is up to the caller to do so after
// this function returns.
// If this is a container with a terminal, we will stream raw. If it is not, we
// will stream with an 8-byte header to multiplex STDOUT and STDERR.
// Returns any errors that occurred, and whether the connection was successfully
// hijacked before that error occurred.
func (r *ConmonOCIRuntime) HTTPAttach(ctr *Container, req *http.Request, w http.ResponseWriter, streams *HTTPAttachStreams, detachKeys *string, cancel <-chan bool, hijackDone chan<- bool, streamAttach, streamLogs bool) (deferredErr error) {
	isTerminal := false
	if ctr.config.Spec.Process != nil {
		isTerminal = ctr.config.Spec.Process.Terminal
	}

	if streams != nil {
		if !streams.Stdin && !streams.Stdout && !streams.Stderr {
			return errors.Wrapf(define.ErrInvalidArg, "must specify at least one stream to attach to")
		}
	}

	attachSock, err := r.AttachSocketPath(ctr)
	if err != nil {
		return err
	}

	var conn *net.UnixConn
	if streamAttach {
		newConn, err := openUnixSocket(attachSock)
		if err != nil {
			return errors.Wrapf(err, "failed to connect to container's attach socket: %v", attachSock)
		}
		conn = newConn
		defer func() {
			if err := conn.Close(); err != nil {
				logrus.Errorf("unable to close container %s attach socket: %q", ctr.ID(), err)
			}
		}()

		logrus.Debugf("Successfully connected to container %s attach socket %s", ctr.ID(), attachSock)
	}

	detachString := ctr.runtime.config.Engine.DetachKeys
	if detachKeys != nil {
		detachString = *detachKeys
	}
	detach, err := processDetachKeys(detachString)
	if err != nil {
		return err
	}

	attachStdout := true
	attachStderr := true
	attachStdin := true
	if streams != nil {
		attachStdout = streams.Stdout
		attachStderr = streams.Stderr
		attachStdin = streams.Stdin
	}

	logrus.Debugf("Going to hijack container %s attach connection", ctr.ID())

	// Alright, let's hijack.
	hijacker, ok := w.(http.Hijacker)
	if !ok {
		return errors.Errorf("unable to hijack connection")
	}

	httpCon, httpBuf, err := hijacker.Hijack()
	if err != nil {
		return errors.Wrapf(err, "error hijacking connection")
	}

	hijackDone <- true

	writeHijackHeader(req, httpBuf)

	// Force a flush after the header is written.
	if err := httpBuf.Flush(); err != nil {
		return errors.Wrapf(err, "error flushing HTTP hijack header")
	}

	defer func() {
		hijackWriteErrorAndClose(deferredErr, ctr.ID(), isTerminal, httpCon, httpBuf)
	}()

	logrus.Debugf("Hijack for container %s attach session done, ready to stream", ctr.ID())

	// TODO: This is gross. Really, really gross.
	// I want to say we should read all the logs into an array before
	// calling this, in container_api.go, but that could take a lot of
	// memory...
	// On the whole, we need to figure out a better way of doing this,
	// though.
	logSize := 0
	if streamLogs {
		logrus.Debugf("Will stream logs for container %s attach session", ctr.ID())

		// Get all logs for the container
		logChan := make(chan *logs.LogLine)
		logOpts := new(logs.LogOptions)
		logOpts.Tail = -1
		logOpts.WaitGroup = new(sync.WaitGroup)
		errChan := make(chan error)
		go func() {
			var err error
			// In non-terminal mode we need to prepend with the
			// stream header.
			logrus.Debugf("Writing logs for container %s to HTTP attach", ctr.ID())
			for logLine := range logChan {
				if !isTerminal {
					device := logLine.Device
					var header []byte
					headerLen := uint32(len(logLine.Msg))
					logSize += len(logLine.Msg)
					switch strings.ToLower(device) {
					case "stdin":
						header = makeHTTPAttachHeader(0, headerLen)
					case "stdout":
						header = makeHTTPAttachHeader(1, headerLen)
					case "stderr":
						header = makeHTTPAttachHeader(2, headerLen)
					default:
						logrus.Errorf("Unknown device for log line: %s", device)
						header = makeHTTPAttachHeader(1, headerLen)
					}
					_, err = httpBuf.Write(header)
					if err != nil {
						break
					}
				}
				_, err = httpBuf.Write([]byte(logLine.Msg))
				if err != nil {
					break
				}
				_, err = httpBuf.Write([]byte("\n"))
				if err != nil {
					break
				}
				err = httpBuf.Flush()
				if err != nil {
					break
				}
			}
			errChan <- err
		}()
		go func() {
			logOpts.WaitGroup.Wait()
			close(logChan)
		}()
		if err := ctr.ReadLog(context.Background(), logOpts, logChan); err != nil {
			return err
		}
		logrus.Debugf("Done reading logs for container %s, %d bytes", ctr.ID(), logSize)
		if err := <-errChan; err != nil {
			return err
		}
	}
	if !streamAttach {
		logrus.Debugf("Done streaming logs for container %s attach, exiting as attach streaming not requested", ctr.ID())
		return nil
	}

	logrus.Debugf("Forwarding attach output for container %s", ctr.ID())

	stdoutChan := make(chan error)
	stdinChan := make(chan error)

	// Handle STDOUT/STDERR
	go func() {
		var err error
		if isTerminal {
			// Hack: return immediately if attachStdout not set to
			// emulate Docker.
			// Basically, when terminal is set, STDERR goes nowhere.
			// Everything does over STDOUT.
			// Therefore, if not attaching STDOUT - we'll never copy
			// anything from here.
			logrus.Debugf("Performing terminal HTTP attach for container %s", ctr.ID())
			if attachStdout {
				err = httpAttachTerminalCopy(conn, httpBuf, ctr.ID())
			}
		} else {
			logrus.Debugf("Performing non-terminal HTTP attach for container %s", ctr.ID())
			err = httpAttachNonTerminalCopy(conn, httpBuf, ctr.ID(), attachStdin, attachStdout, attachStderr)
		}
		stdoutChan <- err
		logrus.Debugf("STDOUT/ERR copy completed")
	}()
	// Next, STDIN. Avoid entirely if attachStdin unset.
	if attachStdin {
		go func() {
			_, err := utils.CopyDetachable(conn, httpBuf, detach)
			logrus.Debugf("STDIN copy completed")
			stdinChan <- err
		}()
	}

	for {
		select {
		case err := <-stdoutChan:
			if err != nil {
				return err
			}

			return nil
		case err := <-stdinChan:
			if err != nil {
				return err
			}
		case <-cancel:
			return nil
		}
	}
}

// isRetryable returns whether the error was caused by a blocked syscall or the
// specified operation on a non blocking file descriptor wasn't ready for completion.
func isRetryable(err error) bool {
	if errno, isErrno := errors.Cause(err).(syscall.Errno); isErrno {
		return errno == syscall.EINTR || errno == syscall.EAGAIN
	}
	return false
}

// openControlFile opens the terminal control file.
func openControlFile(ctr *Container, parentDir string) (*os.File, error) {
	controlPath := filepath.Join(parentDir, "ctl")
	for i := 0; i < 600; i++ {
		controlFile, err := os.OpenFile(controlPath, unix.O_WRONLY|unix.O_NONBLOCK, 0)
		if err == nil {
			return controlFile, err
		}
		if !isRetryable(err) {
			return nil, errors.Wrapf(err, "could not open ctl file for terminal resize for container %s", ctr.ID())
		}
		time.Sleep(time.Second / 10)
	}
	return nil, errors.Errorf("timeout waiting for %q", controlPath)
}

// AttachResize resizes the terminal used by the given container.
func (r *ConmonOCIRuntime) AttachResize(ctr *Container, newSize remotecommand.TerminalSize) error {
	controlFile, err := openControlFile(ctr, ctr.bundlePath())
	if err != nil {
		return err
	}
	defer controlFile.Close()

	logrus.Debugf("Received a resize event for container %s: %+v", ctr.ID(), newSize)
	if _, err = fmt.Fprintf(controlFile, "%d %d %d\n", 1, newSize.Height, newSize.Width); err != nil {
		return errors.Wrapf(err, "failed to write to ctl file to resize terminal")
	}

	return nil
}

// CheckpointContainer checkpoints the given container.
func (r *ConmonOCIRuntime) CheckpointContainer(ctr *Container, options ContainerCheckpointOptions) error {
	if err := label.SetSocketLabel(ctr.ProcessLabel()); err != nil {
		return err
	}
	// imagePath is used by CRIU to store the actual checkpoint files
	imagePath := ctr.CheckpointPath()
	if options.PreCheckPoint {
		imagePath = ctr.PreCheckPointPath()
	}
	// workPath will be used to store dump.log and stats-dump
	workPath := ctr.bundlePath()
	logrus.Debugf("Writing checkpoint to %s", imagePath)
	logrus.Debugf("Writing checkpoint logs to %s", workPath)
	logrus.Debugf("Pre-dump the container %t", options.PreCheckPoint)
	args := []string{}
	args = append(args, r.runtimeFlags...)
	args = append(args, "checkpoint")
	args = append(args, "--image-path")
	args = append(args, imagePath)
	args = append(args, "--work-path")
	args = append(args, workPath)
	if options.KeepRunning {
		args = append(args, "--leave-running")
	}
	if options.TCPEstablished {
		args = append(args, "--tcp-established")
	}
	if !options.PreCheckPoint && options.KeepRunning {
		args = append(args, "--leave-running")
	}
	if options.PreCheckPoint {
		args = append(args, "--pre-dump")
	}
	if !options.PreCheckPoint && options.WithPrevious {
		args = append(args, "--parent-path", ctr.PreCheckPointPath())
	}
	runtimeDir, err := util.GetRuntimeDir()
	if err != nil {
		return err
	}
	if err = os.Setenv("XDG_RUNTIME_DIR", runtimeDir); err != nil {
		return errors.Wrapf(err, "cannot set XDG_RUNTIME_DIR")
	}
	args = append(args, ctr.ID())
	logrus.Debugf("the args to checkpoint: %s %s", r.path, strings.Join(args, " "))
	return utils.ExecCmdWithStdStreams(os.Stdin, os.Stdout, os.Stderr, nil, r.path, args...)
}

func (r *ConmonOCIRuntime) CheckConmonRunning(ctr *Container) (bool, error) {
	if ctr.state.ConmonPID == 0 {
		// If the container is running or paused, assume Conmon is
		// running. We didn't record Conmon PID on some old versions, so
		// that is likely what's going on...
		// Unusual enough that we should print a warning message though.
		if ctr.ensureState(define.ContainerStateRunning, define.ContainerStatePaused) {
			logrus.Warnf("Conmon PID is not set, but container is running!")
			return true, nil
		}
		// Container's not running, so conmon PID being unset is
		// expected. Conmon is not running.
		return false, nil
	}

	// We have a conmon PID. Ping it with signal 0.
	if err := unix.Kill(ctr.state.ConmonPID, 0); err != nil {
		if err == unix.ESRCH {
			return false, nil
		}
		return false, errors.Wrapf(err, "error pinging container %s conmon with signal 0", ctr.ID())
	}
	return true, nil
}

// SupportsCheckpoint checks if the OCI runtime supports checkpointing
// containers.
func (r *ConmonOCIRuntime) SupportsCheckpoint() bool {
	// Check if the runtime implements checkpointing. Currently only
	// runc's checkpoint/restore implementation is supported.
	cmd := exec.Command(r.path, "checkpoint", "--help")
	if err := cmd.Start(); err != nil {
		return false
	}
	if err := cmd.Wait(); err == nil {
		return true
	}
	return false
}

// SupportsJSONErrors checks if the OCI runtime supports JSON-formatted error
// messages.
func (r *ConmonOCIRuntime) SupportsJSONErrors() bool {
	return r.supportsJSON
}

// SupportsNoCgroups checks if the OCI runtime supports running containers
// without cgroups (the --cgroup-manager=disabled flag).
func (r *ConmonOCIRuntime) SupportsNoCgroups() bool {
	return r.supportsNoCgroups
}

// SupportsKVM checks if the OCI runtime supports running containers
// without KVM separation
func (r *ConmonOCIRuntime) SupportsKVM() bool {
	return r.supportsKVM
}

// AttachSocketPath is the path to a single container's attach socket.
func (r *ConmonOCIRuntime) AttachSocketPath(ctr *Container) (string, error) {
	if ctr == nil {
		return "", errors.Wrapf(define.ErrInvalidArg, "must provide a valid container to get attach socket path")
	}

	return filepath.Join(r.socketsDir, ctr.ID(), "attach"), nil
}

// ExitFilePath is the path to a container's exit file.
func (r *ConmonOCIRuntime) ExitFilePath(ctr *Container) (string, error) {
	if ctr == nil {
		return "", errors.Wrapf(define.ErrInvalidArg, "must provide a valid container to get exit file path")
	}
	return filepath.Join(r.exitsDir, ctr.ID()), nil
}

// RuntimeInfo provides information on the runtime.
func (r *ConmonOCIRuntime) RuntimeInfo() (*define.ConmonInfo, *define.OCIRuntimeInfo, error) {
	runtimePackage := packageVersion(r.path)
	conmonPackage := packageVersion(r.conmonPath)
	runtimeVersion, err := r.getOCIRuntimeVersion()
	if err != nil {
		return nil, nil, errors.Wrapf(err, "error getting version of OCI runtime %s", r.name)
	}
	conmonVersion, err := r.getConmonVersion()
	if err != nil {
		return nil, nil, errors.Wrapf(err, "error getting conmon version")
	}

	conmon := define.ConmonInfo{
		Package: conmonPackage,
		Path:    r.conmonPath,
		Version: conmonVersion,
	}
	ocirt := define.OCIRuntimeInfo{
		Name:    r.name,
		Path:    r.path,
		Package: runtimePackage,
		Version: runtimeVersion,
	}
	return &conmon, &ocirt, nil
}

// makeAccessible changes the path permission and each parent directory to have --x--x--x
func makeAccessible(path string, uid, gid int) error {
	for ; path != "/"; path = filepath.Dir(path) {
		st, err := os.Stat(path)
		if err != nil {
			if os.IsNotExist(err) {
				return nil
			}
			return err
		}
		if int(st.Sys().(*syscall.Stat_t).Uid) == uid && int(st.Sys().(*syscall.Stat_t).Gid) == gid {
			continue
		}
		if st.Mode()&0111 != 0111 {
			if err := os.Chmod(path, st.Mode()|0111); err != nil {
				return err
			}
		}
	}
	return nil
}

// Wait for a container which has been sent a signal to stop
func waitContainerStop(ctr *Container, timeout time.Duration) error {
	return waitPidStop(ctr.state.PID, timeout)
}

// Wait for a given PID to stop
func waitPidStop(pid int, timeout time.Duration) error {
	done := make(chan struct{})
	chControl := make(chan struct{})
	go func() {
		for {
			select {
			case <-chControl:
				return
			default:
				if err := unix.Kill(pid, 0); err != nil {
					if err == unix.ESRCH {
						close(done)
						return
					}
					logrus.Errorf("Error pinging PID %d with signal 0: %v", pid, err)
				}
				time.Sleep(100 * time.Millisecond)
			}
		}
	}()
	select {
	case <-done:
		return nil
	case <-time.After(timeout):
		close(chControl)
		return errors.Errorf("given PIDs did not die within timeout")
	}
}

func (r *ConmonOCIRuntime) getLogTag(ctr *Container) (string, error) {
	logTag := ctr.LogTag()
	if logTag == "" {
		return "", nil
	}
	data, err := ctr.inspectLocked(false)
	if err != nil {
		return "", nil
	}
	tmpl, err := template.New("container").Parse(logTag)
	if err != nil {
		return "", errors.Wrapf(err, "template parsing error %s", logTag)
	}
	var b bytes.Buffer
	err = tmpl.Execute(&b, data)
	if err != nil {
		return "", err
	}
	return b.String(), nil
}

// createOCIContainer generates this container's main conmon instance and prepares it for starting
func (r *ConmonOCIRuntime) createOCIContainer(ctr *Container, restoreOptions *ContainerCheckpointOptions) error {
	var stderrBuf bytes.Buffer

	runtimeDir, err := util.GetRuntimeDir()
	if err != nil {
		return err
	}

	parentSyncPipe, childSyncPipe, err := newPipe()
	if err != nil {
		return errors.Wrapf(err, "error creating socket pair")
	}
	defer errorhandling.CloseQuiet(parentSyncPipe)

	childStartPipe, parentStartPipe, err := newPipe()
	if err != nil {
		return errors.Wrapf(err, "error creating socket pair for start pipe")
	}

	defer errorhandling.CloseQuiet(parentStartPipe)

	var ociLog string
	if logrus.GetLevel() != logrus.DebugLevel && r.supportsJSON {
		ociLog = filepath.Join(ctr.state.RunDir, "oci-log")
	}

	logTag, err := r.getLogTag(ctr)
	if err != nil {
		return err
	}

	if ctr.config.CgroupsMode == cgroupSplit {
		if err := utils.MoveUnderCgroupSubtree("supervisor"); err != nil {
			return err
		}
	}

	if ctr.config.SdNotifyMode == define.SdNotifyModeIgnore {
		if err := os.Unsetenv("NOTIFY_SOCKET"); err != nil {
			logrus.Warnf("Error unsetting NOTIFY_SOCKET %v", err)
		}
	}

	args := r.sharedConmonArgs(ctr, ctr.ID(), ctr.bundlePath(), filepath.Join(ctr.state.RunDir, "pidfile"), ctr.LogPath(), r.exitsDir, ociLog, ctr.LogDriver(), logTag)

	if ctr.config.Spec.Process.Terminal {
		args = append(args, "-t")
	} else if ctr.config.Stdin {
		args = append(args, "-i")
	}

	if !r.enableKeyring {
		args = append(args, "--no-new-keyring")
	}
	if ctr.config.ConmonPidFile != "" {
		args = append(args, "--conmon-pidfile", ctr.config.ConmonPidFile)
	}

	if r.noPivot {
		args = append(args, "--no-pivot")
	}

	if len(ctr.config.ExitCommand) > 0 {
		args = append(args, "--exit-command", ctr.config.ExitCommand[0])
		for _, arg := range ctr.config.ExitCommand[1:] {
			args = append(args, []string{"--exit-command-arg", arg}...)
		}
	}

	if ctr.config.PreserveFDs > 0 {
		args = append(args, formatRuntimeOpts("--preserve-fds", fmt.Sprintf("%d", ctr.config.PreserveFDs))...)
	}

	if restoreOptions != nil {
		args = append(args, "--restore", ctr.CheckpointPath())
		if restoreOptions.TCPEstablished {
			args = append(args, "--runtime-opt", "--tcp-established")
		}
	}

	logrus.WithFields(logrus.Fields{
		"args": args,
	}).Debugf("running conmon: %s", r.conmonPath)

	cmd := exec.Command(r.conmonPath, args...)
	cmd.Dir = ctr.bundlePath()
	cmd.SysProcAttr = &syscall.SysProcAttr{
		Setpgid: true,
	}
	// TODO this is probably a really bad idea for some uses
	// Make this configurable
	cmd.Stdin = os.Stdin
	cmd.Stdout = os.Stdout
	cmd.Stderr = os.Stderr
	if ctr.config.Spec.Process.Terminal {
		cmd.Stderr = &stderrBuf
	}

	// 0, 1 and 2 are stdin, stdout and stderr
	conmonEnv, envFiles := r.configureConmonEnv(ctr, runtimeDir)

	var filesToClose []*os.File
	if ctr.config.PreserveFDs > 0 {
		for fd := 3; fd < int(3+ctr.config.PreserveFDs); fd++ {
			f := os.NewFile(uintptr(fd), fmt.Sprintf("fd-%d", fd))
			filesToClose = append(filesToClose, f)
			cmd.ExtraFiles = append(cmd.ExtraFiles, f)
		}
	}

	cmd.Env = r.conmonEnv
	// we don't want to step on users fds they asked to preserve
	// Since 0-2 are used for stdio, start the fds we pass in at preserveFDs+3
	cmd.Env = append(cmd.Env, fmt.Sprintf("_OCI_SYNCPIPE=%d", ctr.config.PreserveFDs+3), fmt.Sprintf("_OCI_STARTPIPE=%d", ctr.config.PreserveFDs+4))
	cmd.Env = append(cmd.Env, conmonEnv...)
	cmd.ExtraFiles = append(cmd.ExtraFiles, childSyncPipe, childStartPipe)
	cmd.ExtraFiles = append(cmd.ExtraFiles, envFiles...)

	if r.reservePorts && !rootless.IsRootless() && !ctr.config.NetMode.IsSlirp4netns() {
		ports, err := bindPorts(ctr.config.PortMappings)
		if err != nil {
			return err
		}

		// Leak the port we bound in the conmon process.  These fd's won't be used
		// by the container and conmon will keep the ports busy so that another
		// process cannot use them.
		cmd.ExtraFiles = append(cmd.ExtraFiles, ports...)
	}

	if ctr.config.NetMode.IsSlirp4netns() || rootless.IsRootless() {
		if ctr.config.PostConfigureNetNS {
			havePortMapping := len(ctr.Config().PortMappings) > 0
			if havePortMapping {
				ctr.rootlessPortSyncR, ctr.rootlessPortSyncW, err = os.Pipe()
				if err != nil {
					return errors.Wrapf(err, "failed to create rootless port sync pipe")
				}
			}
			ctr.rootlessSlirpSyncR, ctr.rootlessSlirpSyncW, err = os.Pipe()
			if err != nil {
				return errors.Wrapf(err, "failed to create rootless network sync pipe")
			}
		} else {
			if ctr.rootlessSlirpSyncR != nil {
				defer errorhandling.CloseQuiet(ctr.rootlessSlirpSyncR)
			}
			if ctr.rootlessSlirpSyncW != nil {
				defer errorhandling.CloseQuiet(ctr.rootlessSlirpSyncW)
			}
		}
		// Leak one end in conmon, the other one will be leaked into slirp4netns
		cmd.ExtraFiles = append(cmd.ExtraFiles, ctr.rootlessSlirpSyncW)

		if ctr.rootlessPortSyncW != nil {
			defer errorhandling.CloseQuiet(ctr.rootlessPortSyncW)
			// Leak one end in conmon, the other one will be leaked into rootlessport
			cmd.ExtraFiles = append(cmd.ExtraFiles, ctr.rootlessPortSyncW)
		}
	}

	err = startCommandGivenSelinux(cmd)
	// regardless of whether we errored or not, we no longer need the children pipes
	childSyncPipe.Close()
	childStartPipe.Close()
	if err != nil {
		return err
	}
	if err := r.moveConmonToCgroupAndSignal(ctr, cmd, parentStartPipe); err != nil {
		return err
	}
	/* Wait for initial setup and fork, and reap child */
	err = cmd.Wait()
	if err != nil {
		return err
	}

	pid, err := readConmonPipeData(parentSyncPipe, ociLog)
	if err != nil {
		if err2 := r.DeleteContainer(ctr); err2 != nil {
			logrus.Errorf("Error removing container %s from runtime after creation failed", ctr.ID())
		}
		return err
	}
	ctr.state.PID = pid

	conmonPID, err := readConmonPidFile(ctr.config.ConmonPidFile)
	if err != nil {
		logrus.Warnf("error reading conmon pid file for container %s: %v", ctr.ID(), err)
	} else if conmonPID > 0 {
		// conmon not having a pid file is a valid state, so don't set it if we don't have it
		logrus.Infof("Got Conmon PID as %d", conmonPID)
		ctr.state.ConmonPID = conmonPID
		if ctr.config.SdNotifyMode != define.SdNotifyModeIgnore {
			if sent, err := daemon.SdNotify(false, fmt.Sprintf("MAINPID=%d", conmonPID)); err != nil {
				logrus.Errorf("Error notifying systemd of Conmon PID: %v", err)
			} else if sent {
				logrus.Debugf("Notify MAINPID sent successfully")
			}
		}
	}

	// These fds were passed down to the runtime.  Close them
	// and not interfere
	for _, f := range filesToClose {
		errorhandling.CloseQuiet(f)
	}

	return nil
}

// prepareProcessExec returns the path of the process.json used in runc exec -p
// caller is responsible to close the returned *os.File if needed.
func prepareProcessExec(c *Container, options *ExecOptions, env []string, sessionID string) (*os.File, error) {
	f, err := ioutil.TempFile(c.execBundlePath(sessionID), "exec-process-")
	if err != nil {
		return nil, err
	}
	pspec := new(spec.Process)
	if err := JSONDeepCopy(c.config.Spec.Process, pspec); err != nil {
		return nil, err
	}
	pspec.SelinuxLabel = c.config.ProcessLabel
	pspec.Args = options.Cmd

	// We need to default this to false else it will inherit terminal as true
	// from the container.
	pspec.Terminal = false
	if options.Terminal {
		pspec.Terminal = true
	}
	if len(env) > 0 {
		pspec.Env = append(pspec.Env, env...)
	}

	if options.Cwd != "" {
		pspec.Cwd = options.Cwd
	}

	var addGroups []string
	var sgids []uint32

	// if the user is empty, we should inherit the user that the container is currently running with
	user := options.User
	if user == "" {
		user = c.config.User
		addGroups = c.config.Groups
	}

	overrides := c.getUserOverrides()
	execUser, err := lookup.GetUserGroupInfo(c.state.Mountpoint, user, overrides)
	if err != nil {
		return nil, err
	}

	if len(addGroups) > 0 {
		sgids, err = lookup.GetContainerGroups(addGroups, c.state.Mountpoint, overrides)
		if err != nil {
			return nil, errors.Wrapf(err, "error looking up supplemental groups for container %s exec session %s", c.ID(), sessionID)
		}
	}

	// If user was set, look it up in the container to get a UID to use on
	// the host
	if user != "" || len(sgids) > 0 {
		if user != "" {
			for _, sgid := range execUser.Sgids {
				sgids = append(sgids, uint32(sgid))
			}
		}
		processUser := spec.User{
			UID:            uint32(execUser.Uid),
			GID:            uint32(execUser.Gid),
			AdditionalGids: sgids,
		}

		pspec.User = processUser
	}

	ctrSpec, err := c.specFromState()
	if err != nil {
		return nil, err
	}

	allCaps := capabilities.AllCapabilities()
	if options.Privileged {
		pspec.Capabilities.Bounding = allCaps
	} else {
		pspec.Capabilities.Bounding = ctrSpec.Process.Capabilities.Bounding
	}
	if execUser.Uid == 0 {
		pspec.Capabilities.Effective = pspec.Capabilities.Bounding
		pspec.Capabilities.Inheritable = pspec.Capabilities.Bounding
		pspec.Capabilities.Permitted = pspec.Capabilities.Bounding
		pspec.Capabilities.Ambient = pspec.Capabilities.Bounding
	} else {
		if user == c.config.User {
			pspec.Capabilities.Effective = ctrSpec.Process.Capabilities.Effective
			pspec.Capabilities.Inheritable = ctrSpec.Process.Capabilities.Effective
			pspec.Capabilities.Permitted = ctrSpec.Process.Capabilities.Effective
			pspec.Capabilities.Ambient = ctrSpec.Process.Capabilities.Effective
		}
	}

	hasHomeSet := false
	for _, s := range pspec.Env {
		if strings.HasPrefix(s, "HOME=") {
			hasHomeSet = true
			break
		}
	}
	if !hasHomeSet {
		pspec.Env = append(pspec.Env, fmt.Sprintf("HOME=%s", execUser.Home))
	}

	processJSON, err := json.Marshal(pspec)
	if err != nil {
		return nil, err
	}

	if err := ioutil.WriteFile(f.Name(), processJSON, 0644); err != nil {
		return nil, err
	}
	return f, nil
}

// configureConmonEnv gets the environment values to add to conmon's exec struct
// TODO this may want to be less hardcoded/more configurable in the future
func (r *ConmonOCIRuntime) configureConmonEnv(ctr *Container, runtimeDir string) ([]string, []*os.File) {
	var env []string
	for _, e := range os.Environ() {
		if strings.HasPrefix(e, "LC_") {
			env = append(env, e)
		}
	}
	env = append(env, fmt.Sprintf("XDG_RUNTIME_DIR=%s", runtimeDir))
	env = append(env, fmt.Sprintf("_CONTAINERS_USERNS_CONFIGURED=%s", os.Getenv("_CONTAINERS_USERNS_CONFIGURED")))
	env = append(env, fmt.Sprintf("_CONTAINERS_ROOTLESS_UID=%s", os.Getenv("_CONTAINERS_ROOTLESS_UID")))
	home := homedir.Get()
	if home != "" {
		env = append(env, fmt.Sprintf("HOME=%s", home))
	}

	extraFiles := make([]*os.File, 0)
	if ctr.config.SdNotifyMode == define.SdNotifyModeContainer {
		if notify, ok := os.LookupEnv("NOTIFY_SOCKET"); ok {
			env = append(env, fmt.Sprintf("NOTIFY_SOCKET=%s", notify))
		}
	}
	if !r.sdNotify {
		if listenfds, ok := os.LookupEnv("LISTEN_FDS"); ok {
			env = append(env, fmt.Sprintf("LISTEN_FDS=%s", listenfds), "LISTEN_PID=1")
			fds := activation.Files(false)
			extraFiles = append(extraFiles, fds...)
		}
	} else {
		logrus.Debug("disabling SD notify")
	}
	return env, extraFiles
}

// sharedConmonArgs takes common arguments for exec and create/restore and formats them for the conmon CLI
func (r *ConmonOCIRuntime) sharedConmonArgs(ctr *Container, cuuid, bundlePath, pidPath, logPath, exitDir, ociLogPath, logDriver, logTag string) []string {
	// set the conmon API version to be able to use the correct sync struct keys
	args := []string{
		"--api-version", "1",
		"-c", ctr.ID(),
		"-u", cuuid,
		"-r", r.path,
		"-b", bundlePath,
		"-p", pidPath,
		"-n", ctr.Name(),
		"--exit-dir", exitDir,
		"--socket-dir-path", r.socketsDir,
	}
	if len(r.runtimeFlags) > 0 {
		rFlags := []string{}
		for _, arg := range r.runtimeFlags {
			rFlags = append(rFlags, "--runtime-arg", arg)
		}
		args = append(args, rFlags...)
	}

	if ctr.CgroupManager() == config.SystemdCgroupsManager && !ctr.config.NoCgroups && ctr.config.CgroupsMode != cgroupSplit {
		args = append(args, "-s")
	}

	var logDriverArg string
	switch logDriver {
	case define.JournaldLogging:
		logDriverArg = define.JournaldLogging
	case define.NoLogging:
		logDriverArg = define.NoLogging
	case define.JSONLogging:
		fallthrough
	//lint:ignore ST1015 the default case has to be here
	default: //nolint-stylecheck
		// No case here should happen except JSONLogging, but keep this here in case the options are extended
		logrus.Errorf("%s logging specified but not supported. Choosing k8s-file logging instead", ctr.LogDriver())
		fallthrough
	case "":
		// to get here, either a user would specify `--log-driver ""`, or this came from another place in libpod
		// since the former case is obscure, and the latter case isn't an error, let's silently fallthrough
		fallthrough
	case define.KubernetesLogging:
		logDriverArg = fmt.Sprintf("%s:%s", define.KubernetesLogging, logPath)
	}

	args = append(args, "-l", logDriverArg)
	logLevel := logrus.GetLevel()
	args = append(args, "--log-level", logLevel.String())

	if logLevel == logrus.DebugLevel {
		logrus.Debugf("%s messages will be logged to syslog", r.conmonPath)
		args = append(args, "--syslog")
	}

	size := r.logSizeMax
	if ctr.config.LogSize > 0 {
		size = ctr.config.LogSize
	}
	if size > 0 {
		args = append(args, "--log-size-max", fmt.Sprintf("%v", size))
	}

	if ociLogPath != "" {
		args = append(args, "--runtime-arg", "--log-format=json", "--runtime-arg", "--log", fmt.Sprintf("--runtime-arg=%s", ociLogPath))
	}
	if logTag != "" {
		args = append(args, "--log-tag", logTag)
	}
	if ctr.config.NoCgroups {
		logrus.Debugf("Running with no CGroups")
		args = append(args, "--runtime-arg", "--cgroup-manager", "--runtime-arg", "disabled")
	}
	return args
}

// startCommandGivenSelinux starts a container ensuring to set the labels of
// the process to make sure SELinux doesn't block conmon communication, if SELinux is enabled
func startCommandGivenSelinux(cmd *exec.Cmd) error {
	if !selinux.GetEnabled() {
		return cmd.Start()
	}
	// Set the label of the conmon process to be level :s0
	// This will allow the container processes to talk to fifo-files
	// passed into the container by conmon
	var (
		plabel string
		con    selinux.Context
		err    error
	)
	plabel, err = selinux.CurrentLabel()
	if err != nil {
		return errors.Wrapf(err, "failed to get current SELinux label")
	}

	con, err = selinux.NewContext(plabel)
	if err != nil {
		return errors.Wrapf(err, "failed to get new context from SELinux label")
	}

	runtime.LockOSThread()
	if con["level"] != "s0" && con["level"] != "" {
		con["level"] = "s0"
		if err = label.SetProcessLabel(con.Get()); err != nil {
			runtime.UnlockOSThread()
			return err
		}
	}
	err = cmd.Start()
	// Ignore error returned from SetProcessLabel("") call,
	// can't recover.
	if labelErr := label.SetProcessLabel(""); labelErr != nil {
		logrus.Errorf("unable to set process label: %q", err)
	}
	runtime.UnlockOSThread()
	return err
}

// moveConmonToCgroupAndSignal gets a container's cgroupParent and moves the conmon process to that cgroup
// it then signals for conmon to start by sending nonce data down the start fd
func (r *ConmonOCIRuntime) moveConmonToCgroupAndSignal(ctr *Container, cmd *exec.Cmd, startFd *os.File) error {
	mustCreateCgroup := true

	if ctr.config.NoCgroups {
		mustCreateCgroup = false
	}

	// If cgroup creation is disabled - just signal.
	switch ctr.config.CgroupsMode {
	case "disabled", "no-conmon", cgroupSplit:
		mustCreateCgroup = false
	}

	// $INVOCATION_ID is set by systemd when running as a service.
	if os.Getenv("INVOCATION_ID") != "" {
		mustCreateCgroup = false
	}

	if mustCreateCgroup {
		// Usually rootless users are not allowed to configure cgroupfs.
		// There are cases though, where it is allowed, e.g. if the cgroup
		// is manually configured and chowned).  Avoid detecting all
		// such cases and simply use a lower log level.
		logLevel := logrus.WarnLevel
		if rootless.IsRootless() {
			logLevel = logrus.InfoLevel
		}
		// TODO: This should be a switch - we are not guaranteed that
		// there are only 2 valid cgroup managers
		cgroupParent := ctr.CgroupParent()
		if ctr.CgroupManager() == config.SystemdCgroupsManager {
			unitName := createUnitName("libpod-conmon", ctr.ID())

			realCgroupParent := cgroupParent
			splitParent := strings.Split(cgroupParent, "/")
			if strings.HasSuffix(cgroupParent, ".slice") && len(splitParent) > 1 {
				realCgroupParent = splitParent[len(splitParent)-1]
			}

			logrus.Infof("Running conmon under slice %s and unitName %s", realCgroupParent, unitName)
			if err := utils.RunUnderSystemdScope(cmd.Process.Pid, realCgroupParent, unitName); err != nil {
				logrus.StandardLogger().Logf(logLevel, "Failed to add conmon to systemd sandbox cgroup: %v", err)
			}
		} else {
			cgroupPath := filepath.Join(ctr.config.CgroupParent, "conmon")
			control, err := cgroups.New(cgroupPath, &spec.LinuxResources{})
			if err != nil {
				logrus.StandardLogger().Logf(logLevel, "Failed to add conmon to cgroupfs sandbox cgroup: %v", err)
			} else if err := control.AddPid(cmd.Process.Pid); err != nil {
				// we need to remove this defer and delete the cgroup once conmon exits
				// maybe need a conmon monitor?
				logrus.StandardLogger().Logf(logLevel, "Failed to add conmon to cgroupfs sandbox cgroup: %v", err)
			}
		}
	}

	/* We set the cgroup, now the child can start creating children */
	if err := writeConmonPipeData(startFd); err != nil {
		return err
	}
	return nil
}

// newPipe creates a unix socket pair for communication.
// Returns two files - first is parent, second is child.
func newPipe() (*os.File, *os.File, error) {
	fds, err := unix.Socketpair(unix.AF_LOCAL, unix.SOCK_SEQPACKET|unix.SOCK_CLOEXEC, 0)
	if err != nil {
		return nil, nil, err
	}
	return os.NewFile(uintptr(fds[1]), "parent"), os.NewFile(uintptr(fds[0]), "child"), nil
}

// readConmonPidFile attempts to read conmon's pid from its pid file
func readConmonPidFile(pidFile string) (int, error) {
	// Let's try reading the Conmon pid at the same time.
	if pidFile != "" {
		contents, err := ioutil.ReadFile(pidFile)
		if err != nil {
			return -1, err
		}
		// Convert it to an int
		conmonPID, err := strconv.Atoi(string(contents))
		if err != nil {
			return -1, err
		}
		return conmonPID, nil
	}
	return 0, nil
}

// readConmonPipeData attempts to read a syncInfo struct from the pipe
func readConmonPipeData(pipe *os.File, ociLog string) (int, error) {
	// syncInfo is used to return data from monitor process to daemon
	type syncInfo struct {
		Data    int    `json:"data"`
		Message string `json:"message,omitempty"`
	}

	// Wait to get container pid from conmon
	type syncStruct struct {
		si  *syncInfo
		err error
	}
	ch := make(chan syncStruct)
	go func() {
		var si *syncInfo
		rdr := bufio.NewReader(pipe)
		b, err := rdr.ReadBytes('\n')
		if err != nil {
			ch <- syncStruct{err: err}
		}
		if err := json.Unmarshal(b, &si); err != nil {
			ch <- syncStruct{err: err}
			return
		}
		ch <- syncStruct{si: si}
	}()

	data := -1
	select {
	case ss := <-ch:
		if ss.err != nil {
			if ociLog != "" {
				ociLogData, err := ioutil.ReadFile(ociLog)
				if err == nil {
					var ociErr ociError
					if err := json.Unmarshal(ociLogData, &ociErr); err == nil {
						return -1, getOCIRuntimeError(ociErr.Msg)
					}
				}
			}
			return -1, errors.Wrapf(ss.err, "container create failed (no logs from conmon)")
		}
		logrus.Debugf("Received: %d", ss.si.Data)
		if ss.si.Data < 0 {
			if ociLog != "" {
				ociLogData, err := ioutil.ReadFile(ociLog)
				if err == nil {
					var ociErr ociError
					if err := json.Unmarshal(ociLogData, &ociErr); err == nil {
						return ss.si.Data, getOCIRuntimeError(ociErr.Msg)
					}
				}
			}
			// If we failed to parse the JSON errors, then print the output as it is
			if ss.si.Message != "" {
				return ss.si.Data, getOCIRuntimeError(ss.si.Message)
			}
			return ss.si.Data, errors.Wrapf(define.ErrInternal, "container create failed")
		}
		data = ss.si.Data
	case <-time.After(define.ContainerCreateTimeout):
		return -1, errors.Wrapf(define.ErrInternal, "container creation timeout")
	}
	return data, nil
}

// writeConmonPipeData writes nonce data to a pipe
func writeConmonPipeData(pipe *os.File) error {
	someData := []byte{0}
	_, err := pipe.Write(someData)
	return err
}

// formatRuntimeOpts prepends opts passed to it with --runtime-opt for passing to conmon
func formatRuntimeOpts(opts ...string) []string {
	args := make([]string, 0, len(opts)*2)
	for _, o := range opts {
		args = append(args, "--runtime-opt", o)
	}
	return args
}

// getConmonVersion returns a string representation of the conmon version.
func (r *ConmonOCIRuntime) getConmonVersion() (string, error) {
	output, err := utils.ExecCmd(r.conmonPath, "--version")
	if err != nil {
		return "", err
	}
	return strings.TrimSuffix(strings.Replace(output, "\n", ", ", 1), "\n"), nil
}

// getOCIRuntimeVersion returns a string representation of the OCI runtime's
// version.
func (r *ConmonOCIRuntime) getOCIRuntimeVersion() (string, error) {
	output, err := utils.ExecCmd(r.path, "--version")
	if err != nil {
		return "", err
	}
	return strings.TrimSuffix(output, "\n"), nil
}

// Copy data from container to HTTP connection, for terminal attach.
// Container is the container's attach socket connection, http is a buffer for
// the HTTP connection. cid is the ID of the container the attach session is
// running for (used solely for error messages).
func httpAttachTerminalCopy(container *net.UnixConn, http *bufio.ReadWriter, cid string) error {
	buf := make([]byte, bufferSize)
	for {
		numR, err := container.Read(buf)
		logrus.Debugf("Read fd(%d) %d/%d bytes for container %s", int(buf[0]), numR, len(buf), cid)

		if numR > 0 {
			switch buf[0] {
			case AttachPipeStdout:
				// Do nothing
			default:
				logrus.Errorf("Received unexpected attach type %+d, discarding %d bytes", buf[0], numR)
				continue
			}

			numW, err2 := http.Write(buf[1:numR])
			if err2 != nil {
				if err != nil {
					logrus.Errorf("Error reading container %s STDOUT: %v", cid, err)
				}
				return err2
			} else if numW+1 != numR {
				return io.ErrShortWrite
			}
			// We need to force the buffer to write immediately, so
			// there isn't a delay on the terminal side.
			if err2 := http.Flush(); err2 != nil {
				if err != nil {
					logrus.Errorf("Error reading container %s STDOUT: %v", cid, err)
				}
				return err2
			}
		}
		if err != nil {
			if err == io.EOF {
				return nil
			}
			return err
		}
	}
}

// Copy data from a container to an HTTP connection, for non-terminal attach.
// Appends a header to multiplex input.
func httpAttachNonTerminalCopy(container *net.UnixConn, http *bufio.ReadWriter, cid string, stdin, stdout, stderr bool) error {
	buf := make([]byte, bufferSize)
	for {
		numR, err := container.Read(buf)
		if numR > 0 {
			var headerBuf []byte

			// Subtract 1 because we strip the first byte (used for
			// multiplexing by Conmon).
			headerLen := uint32(numR - 1)
			// Practically speaking, we could make this buf[0] - 1,
			// but we need to validate it anyways...
			switch buf[0] {
			case AttachPipeStdin:
				headerBuf = makeHTTPAttachHeader(0, headerLen)
				if !stdin {
					continue
				}
			case AttachPipeStdout:
				if !stdout {
					continue
				}
				headerBuf = makeHTTPAttachHeader(1, headerLen)
			case AttachPipeStderr:
				if !stderr {
					continue
				}
				headerBuf = makeHTTPAttachHeader(2, headerLen)
			default:
				logrus.Errorf("Received unexpected attach type %+d, discarding %d bytes", buf[0], numR)
				continue
			}

			numH, err2 := http.Write(headerBuf)
			if err2 != nil {
				if err != nil {
					logrus.Errorf("Error reading container %s standard streams: %v", cid, err)
				}

				return err2
			}
			// Hardcoding header length is pretty gross, but
			// fast. Should be safe, as this is a fixed part
			// of the protocol.
			if numH != 8 {
				if err != nil {
					logrus.Errorf("Error reading container %s standard streams: %v", cid, err)
				}

				return io.ErrShortWrite
			}

			numW, err2 := http.Write(buf[1:numR])
			if err2 != nil {
				if err != nil {
					logrus.Errorf("Error reading container %s standard streams: %v", cid, err)
				}

				return err2
			} else if numW+1 != numR {
				if err != nil {
					logrus.Errorf("Error reading container %s standard streams: %v", cid, err)
				}

				return io.ErrShortWrite
			}
			// We need to force the buffer to write immediately, so
			// there isn't a delay on the terminal side.
			if err2 := http.Flush(); err2 != nil {
				if err != nil {
					logrus.Errorf("Error reading container %s STDOUT: %v", cid, err)
				}
				return err2
			}
		}
		if err != nil {
			if err == io.EOF {
				return nil
			}

			return err
		}
	}
}